CN108229718B - Information prediction method and device - Google Patents

Information prediction method and device Download PDF

Info

Publication number
CN108229718B
CN108229718B CN201611199287.0A CN201611199287A CN108229718B CN 108229718 B CN108229718 B CN 108229718B CN 201611199287 A CN201611199287 A CN 201611199287A CN 108229718 B CN108229718 B CN 108229718B
Authority
CN
China
Prior art keywords
prediction
irt
question
probability
information
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201611199287.0A
Other languages
Chinese (zh)
Other versions
CN108229718A (en
Inventor
刘源
李历
高钰舒
张凯磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing ByteDance Network Technology Co Ltd
Original Assignee
Beijing ByteDance Network Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing ByteDance Network Technology Co Ltd filed Critical Beijing ByteDance Network Technology Co Ltd
Priority to CN201611199287.0A priority Critical patent/CN108229718B/en
Publication of CN108229718A publication Critical patent/CN108229718A/en
Application granted granted Critical
Publication of CN108229718B publication Critical patent/CN108229718B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/20—Education
    • G06Q50/205—Education administration or guidance

Abstract

The embodiment of the invention discloses an information prediction method and device. The method comprises the following steps: acquiring historical answer information of a current answer user about a target question bank and learning capacity information of the current answer user; determining candidate questions from a target question bank; predicting the probability of the current answer to the candidate question based on the IRT prediction model and the learning ability information to obtain an IRT prediction result; predicting the probability of the current answer to the candidate question based on the DKT prediction model and the historical answer information to obtain a DKT prediction result; and combining the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question. By adopting the technical scheme, the embodiment of the invention synthesizes the prediction results of the two prediction models to predict the probability of the candidate questions made by the answerer, thereby effectively improving the prediction accuracy.

Description

Information prediction method and device
Technical Field
The embodiment of the invention relates to the technical field of information processing, in particular to an information prediction method and device.
Background
With the wide application of computer technology in the field of education, adaptive testing, adaptive learning and the like are increasingly receiving attention of people. The self-adaptive learning system aims to provide a student independent learning platform which collects problem solving information of students, evaluates problem solving capacity of the students in real time through technical means, analyzes learning paths most suitable for the students to master subjects, and integrates and updates problem database data. The self-adaptive learning system has the functions of reasonably optimizing the learning schedule of students, mobilizing the learning enthusiasm of the students, assisting teachers to improve teaching efficiency and solving the problem of uneven distribution of education resources and the like.
The core of the adaptive learning lies in how to effectively evaluate the problem solving information of students and arrange corresponding learning paths through a computer. The study on student test evaluation problems dates back to the Classic Test Theory (CTT) proposed in the 30 s of the twentieth century, which considers the student problem solving results as some linear fit of student ability plus random noise, which contributes greatly to the Theory and practice of both the psychological and educational measures. However, as the times develop, the knowledge contents learned by students become rich and diversified, and the application and development of the CCT theory are limited by the standardization requirements of the CCT theory on test question sets and the difficulty in repetitive implementation of the randomization technology, and the CCT theory cannot meet increasingly diversified teaching modes and daily learning evaluation. Therefore, new theories are emerging, such as Bayesian Knowledge Tracking (BKT) models, Item Response Theory (IRT), and Deep learning Knowledge tracking (DKT).
Although the above models have been applied and popularized to different degrees, the above models still have shortcomings in terms of execution efficiency and accuracy of prediction results.
Disclosure of Invention
The embodiment of the invention aims to provide an information prediction method and an information prediction device, so that the existing scheme for predicting the answer information of an answerer is optimized, and the prediction accuracy is improved.
In one aspect, an embodiment of the present invention provides an information prediction method, including:
acquiring historical answer information of a current answer user about a target question bank and learning capacity information of the current answer user;
determining candidate questions from the target question bank;
predicting the probability of the current answerer for the candidate question based on a project reflection theory IRT prediction model and the learning ability information to obtain an IRT prediction result;
predicting the probability of the current answerer for the candidate question based on a deep learning knowledge tracking DKT prediction model and the historical answer information to obtain a DKT prediction result;
and combining the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
In another aspect, an embodiment of the present invention provides an information prediction apparatus, including:
the information acquisition module is used for acquiring historical answer information of a current answer user about a target question bank and learning capacity information of the current answer user;
the candidate question determining module is used for determining candidate questions from the target question bank;
the first prediction module is used for predicting the probability of the current answerer for the candidate question based on a project reflection theory IRT prediction model and the learning capability information to obtain an IRT prediction result;
the second prediction module is used for predicting the probability of the current answer to the candidate question based on a deep learning knowledge tracking DKT prediction model and the historical answer information to obtain a DKT prediction result;
and the prediction result processing module is used for carrying out merging processing on the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
The information prediction scheme provided by the embodiment of the invention predicts the probability of the candidate questions based on the IRT prediction model and the learning ability information of the current answerer for the candidate questions in the target question bank to obtain the IRT prediction result, predicts the probability of the candidate questions based on the DKT prediction model and the historical answer information of the current answerer to obtain the DKT prediction result, and finally performs combination processing on the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer for the candidate questions. By adopting the technical scheme, the prediction results of the two prediction models are integrated to predict the probability of the candidate questions of the answerer, so that the prediction accuracy can be effectively improved.
Drawings
Fig. 1 is a schematic flowchart of an information prediction method according to an embodiment of the present invention;
fig. 2 is a schematic flowchart of an information prediction method according to a second embodiment of the present invention;
FIG. 3 is a schematic diagram of a topic pushing process according to a second embodiment of the present invention;
FIG. 4a is a first comparison diagram of a technical solution according to an embodiment of the present invention and a prior art solution provided by an embodiment of the present invention;
FIG. 4b is a second comparison between the technical solution of the embodiment of the present invention and the prior art;
fig. 5 is a block diagram of an information prediction apparatus according to a third embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further explained by the specific implementation mode in combination with the attached drawings. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting of the invention. It should be further noted that, for the convenience of description, only some of the structures related to the present invention are shown in the drawings, not all of the structures.
Before discussing exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although a flowchart may describe the steps as a sequential process, many of the steps can be performed in parallel, concurrently or simultaneously. In addition, the order of the steps may be rearranged. The process may be terminated when its operations are completed, but may have additional steps not included in the figure. The processes may correspond to methods, functions, procedures, subroutines, and the like.
To facilitate an understanding of the embodiments of the present invention, a brief description of IRT and DKT follows.
First, the project response theory IRT is also called potential idiom theory or project characteristic curve theory, which is an estimation of the ability of the answerer and links a certain response probability (such as answer-to-probability or wrong-answer probability) of the examinee to a single test project (question) with a certain idiom (such as question discrimination and question difficulty). The characteristic curve contains test question parameters describing the characteristics of the test questions and potential traits or capability parameters describing the characteristics of the respondents. At present, the most widely applied IRT model is a model represented by a logistic model proposed by bahn baum, and according to the difference of the number of parameters, the feature functions can be divided into a single-parameter IRT model, a two-parameter IRT model and a three-parameter IRT model.
The DKT model is a neural network-based model proposed by a learner at the university of Stanford (Stanford) in 2015, can predict the problem-making conditions (such as right or wrong) of students, and is applicable to a self-adaptive learning system. However, the theory and practice of the model are not mature, and the prediction precision and accuracy need to be improved.
The existing IRT model and DKT model each have deficiencies. Because the IRT model uses the maximum posterior probability estimation, certain assumptions are made on prior information of students and subjects, and the evaluation effect is reduced if the assumptions are not good. In addition, the computational efficiency of the complex IRT model is not in direct proportion to the brought yield on the precision, and the marginal cost is too high. The requirement of the existing IRT model on the uniformity of a student problem solving template is too strict, and a method for relevant processing of missing data (nonignorable missing data) is not flexible and efficient. On the other hand, the DKT model is not mature in theory and practice at present, the precision is not improved compared with a complex IRT model, the final given result is only the judgment of the correct problem solving of the student, and information on the learning ability and the problem difficulty attribute of the student is not systematically given.
Example one
Fig. 1 is a schematic flowchart of an information prediction method according to an embodiment of the present invention, where the method may be executed by an information prediction apparatus, where the apparatus may be implemented by software and/or hardware, and may be generally integrated in a terminal in an adaptive learning system, where the terminal may be a terminal such as a personal computer or a server, or a mobile terminal such as a tablet computer or a smart phone, and the embodiment of the present invention is not limited in particular. As shown in fig. 1, the method includes:
and step 110, obtaining historical answering information of the current answering person about the target question bank and learning ability information of the current answering person.
In this embodiment, the target question bank can be selected according to actual requirements. Specifically, the target question bank can be selected according to the identity information input by the current answerer, or the answerer can independently select the target question bank. For example, if the identity information is a student in junior middle school grade, the target question bank may be a question bank of a subject in junior middle school grade; if the identity information is a driver license examiner in the area B, the target subject library can be a subject library of a certain driver test in the area B. Illustratively, the historical answer information may include information such as the number of completed answers, the questions to be answered, and the answer status (e.g., right or wrong) of the questions. The learning ability information may specifically be the learning ability θ in the IRT model, that is, the θ value corresponding to the current answerer.
Step 120, candidate topics are determined from the target topic library.
The embodiment does not limit the determination method of the candidate topics in this step, and the candidate topics may be selected randomly or according to knowledge points.
And step 130, predicting the probability of the current answerer for the candidate question based on the IRT prediction model and the learning ability information to obtain an IRT prediction result.
in the embodiment of the present invention, taking a classical two-parameter IRT model as an example, let θ (theta) be the learning ability of an answerer, α (differentiation) and β (differentiation) be the discrimination and difficulty (coefficient) of a question, respectively, and then the probability of the right of the answerer of the question is:
Figure BDA0001188718000000061
it should be noted that, in the single-parameter IRT model, α is generally replaced by a fixed D (D takes a value of 1.7).
In order to estimate the question information of the questions in the target question bank, the discrimination and difficulty of each question in the question bank need to be estimated according to the answer information of the answers of the questions, and an IRT prediction model is further established.
The establishing process of the IRT prediction model is not particularly limited in this embodiment, and preferably, the IRT prediction model may be established by the following first and/or second manners.
Method for constructing IRT prediction model based on MML (Maximum likelihood of Marginal) estimation method
Acquiring answer information samples of a preset number of answer users about the target question bank before predicting the probability of the current answer user for the candidate question based on the IRT model and the learning ability information and obtaining an IRT prediction result; determining the frequency weight of the corresponding question in the target question bank which is paired by all the answering persons with preset learning capacity based on the assumption that the answering condition of the question made by each answering person is an independent event; substituting the frequency weight into a preset estimation model to obtain a target estimation model, wherein the preset estimation model is a maximum posterior probability estimation model based on an IRT (inverse Fourier transform) model and a Marginal Maximum Likelihood (MML) estimation method; training the target estimation model by using an MML estimation method according to the answer sample information to estimate first question information of each question in the target question bank, wherein the first question information comprises a first discrimination and a first difficulty; and establishing a first IRT prediction model according to the trained target estimation model.
In the above manner, the answer information sample may include information such as the number of answers, the questions to be answered, and the answer situation (e.g., right or wrong) and the like.
Specifically, given an answerer i and a question j in the question bank, the following maximum posterior probability estimate needs to be maximized:
Figure BDA0001188718000000071
wherein i represents an answerer, j represents a question, and Xi,jshows the wrong question setting condition, alpha, of the question i about the question jjindicates the degree of distinction, β, of the topic jjDifficulty coefficient, θ, representing topic jiThe learning ability of the answerer i is expressed,P(Xi,j|αj,βj，θi) The probability of the question maker i making correct/wrong question j is shown,
Figure BDA0001188718000000072
denotes thetaiSatisfy the mean value of
Figure BDA0001188718000000073
Variance of
Figure BDA0001188718000000074
The normal distribution of (c),
Figure BDA0001188718000000075
is represented by betajSatisfy the mean value of
Figure BDA0001188718000000076
Variance of
Figure BDA0001188718000000077
The normal distribution of (c),
Figure BDA0001188718000000078
represents ln alphajSatisfy the mean value of
Figure BDA0001188718000000079
Variance of
Figure BDA00011887180000000710
Is normally distributed. The above formula is written as a maximum a posteriori probability estimation model based on the IRT model.
Common methods for solving the optimization problem include the above-mentioned JML estimation method, MML estimation method, MCMC method, and the like, and the embodiment of the present invention employs the MML estimation method that is widely applied. In the method, a preset estimation model, namely a maximum posterior probability estimation model based on an IRT model and an MML estimation method, can be obtained by assuming that the learning capacity theta meets a certain prior distribution. E.g. thetaiSatisfy the mean value of
Figure BDA00011887180000000711
Variance of
Figure BDA00011887180000000712
Is predetermined prior normal distribution. For convenience of explanation, it is assumed here that θiSatisfying a prior normal distribution with a mean of 0 and a variance of 1, i.e. θiSatisfy the requirement of
Figure BDA00011887180000000713
Based on this assumption, the above IRT model-based maximum a posteriori probability estimation model can be converted into:
Figure BDA00011887180000000714
wherein the content of the first and second substances,
Figure BDA00011887180000000715
represents that the topic j is represented by all learning abilities as thetaiThe frequency of right or wrong, wherein ItemjWhen 1 represents a pair, ItemjWhen 0, it indicates an error.
Further, Gaussian-Hermite integral formula is introduced to simulate the ability sampling of students, and the learning ability of the students is assumed to meet the requirement
Figure BDA0001188718000000081
Normal distribution, the frequency of the student ability theta being x is
Figure BDA0001188718000000082
Wherein Hn-1(x) Is an n-1 order Hermite polynomial, and the learning ability sampling point x is taken as a polynomial Hn-1(x) All possible zeros.
To obtain
Figure BDA0001188718000000083
That the learning ability of the topic j is theta is obtainediThe frequency of right-making of the answerer of (1), conventionalThe idea of analyzing the frequency of various answering conditions of questions is to provide a fixed question group for an answerer, reversely deducing the frequency of the answering conditions of a single question based on the frequency of the various answering conditions of the fixed question group, wherein when the number of the questions is large, the total number of possible patterns of the question group is exponentially increased, the efficiency and the accuracy of the traditional mode are extremely reduced, the requirement on the performance of hardware equipment is extremely high, even the requirement is difficult to meet, and the MML method is not feasible to be applied to the self-adaptive answering environment.
Therefore, the embodiment of the invention abandons the traditional thought of reversely deducing the frequency of the answer condition of a single question based on the frequency of the various answer conditions of the fixed question group when the frequency of the various answer conditions of the question appears, innovatively proposes that certain independence exists on the assumption that the probability of the various answer conditions of the question made by each answerer appears, namely the answer condition of the question made by each answerer can be approximately regarded as an independent event, adopts a processing mode directly starting from the frequency of the various answer conditions of the single question, namely, on the assumption that the answer condition of the question made by each answerer is the independent event, and determines the frequency weight of the corresponding question in the target question bank to be paired by all answerers with preset learning capacity. The determining the frequency weight of the corresponding question in the target question bank being paired by all the respondents with the preset learning ability in the embodiment may specifically refer to: and respectively determining the frequency weight of the current question to be paired by all the respondents with preset learning ability for each question in the target question bank.
Specifically, the frequency weight calculation method is as follows:
Figure BDA0001188718000000091
wherein the content of the first and second substances,
Figure BDA0001188718000000092
represents that the topic j is represented by all learning abilities as thetaiThe frequency of the right-to-right of the answerer,
Figure BDA0001188718000000093
presentation ItemjStatistical frequency of pairings, P (Item)j＝1|θi) Represents that the learning ability in the case where the topic j is paired is θiThe posterior probability of occurrence of the student, P (theta)i) A priori assumptions representing the learning abilities of all answerers.
And then substituting the frequency weight into a preset estimation model to obtain a target estimation model, training the target estimation model by using an MML (multi-media distance) estimation method according to the answer sample information to estimate the first question information of each question in the target question bank, and establishing a first IRT (inter-range transform) prediction model according to the trained target estimation model.
The first IRT prediction model solves the problem that the efficiency and the accuracy of the MML estimation method based on the IRT model are sharply reduced or even infeasible when a large number of questions are included in a question group, so that the MML estimation method based on the IRT model can be better applied to a self-adaptive learning environment, and the efficiency and the estimation accuracy of the model are effectively improved.
Second, constructing IRT prediction model based on variational inference method
Acquiring answer information samples of a preset number of answerers about the target question bank, and constructing a Bayesian network model by taking the learning capacity of the answerers, the discrimination of questions and the difficulty of the questions in the IRT model as parameters to be estimated, wherein the parameters to be estimated meet preset prior distribution containing hyper-parameters; determining a variation distribution function corresponding to a target function by adopting a variation inference method, and estimating the hyper-parameter based on the Bayesian network model and the answer information sample by taking the minimum degree of closeness of the target function and the variation distribution function as a principle to obtain a parameter value of the hyper-parameter, wherein the target function is a posterior estimation function about the parameter to be estimated based on the answer information sample; updating the variation distribution function according to the obtained parameter values of the super parameters; sampling the parameter to be estimated based on the updated variational distribution function to obtain the estimation of the parameter to be estimated; and establishing a second IRT prediction model according to the estimation result of the parameter to be estimated, wherein the estimation result comprises second topic information of each topic in the target topic library, and the second topic information comprises second discrimination and second difficulty.
In a general bayesian network model based on an IRT model, prior information of students is often determined in advance, for example, it is generally assumed that the capability distributions of all students conform to a normal distribution N (0,1), which brings a priori solidity to the estimation of the model, and even though parameters of the normal distribution can be selected, the tuning process makes the whole estimation process to be executed again, which seriously affects the execution efficiency of the estimation model. Therefore, in the embodiment of the invention, by introducing hyper-parameters (hyper-parameters), the parameters to be estimated meet a certain prior hypothesis distribution family, thereby weakening the analysis error brought by parameter error estimation.
Preferably, the learning ability of the answerer and the difficulty of the questions satisfy a normal distribution with a mean and/or a variance as a super-parameter, and the discrimination of the questions satisfies a log-normal distribution with a mean and/or a variance as a super-parameter. The normal distribution in which the variance or even the mean is to be determined is a series of normal distribution functions, which can be referred to as a family of normal distribution functions.
in the embodiment of the invention, it is assumed that α, β, and θ respectively satisfy the super-parameter distribution as follows:
Figure BDA0001188718000000101
wherein, tauθCan satisfy uniform distribution, tau, in the interval of (0,100)αCan satisfy uniform distribution, tau, in the interval of (0,100)βA uniform distribution within the interval (0,100) can be satisfied. It is understood that 100 is a freely settable constant, but may be any other value, so that the variance of any empirical parameter will not exceed this range.
When a general Bayesian network model is used for estimation, an MCMC method can be used for sampling and carrying out integral summation on a priori assumption, but when the Bayesian network model is complex (such as a large number of students or subjects), the efficiency of the MCMC sampler is very slow, and the execution efficiency of the model is influenced. In the embodiment, the variational inference method can well expand the complexity of the model, improve the sampling speed and further improve the execution efficiency of the model.
in particular, let Z be the parameter set to be estimated, Z ═ α, β, θ, α be the degree of distinction of the topic, β be the difficulty of the topic, and θ be the learning ability of the answererα、τβAnd τθThe undetermined prior distribution function can obtain:
p(Z|X)≈q(Z)
based on the principle that the closeness degree of the objective function and the variation distribution function is minimum, namely the objective is to find the objective variation distribution function q corresponding to p (Z | X)*(Z) makes p (Z | X) and q (Z) most nearly equal, and therefore, q can be obtained*(Z) is a distribution that satisfies the minimum value of:
Figure BDA0001188718000000111
the above formula is a defined formula regarding KL divergence (Kullback-Leibler divergence).
Due to the fact that
Figure BDA0001188718000000112
Figure BDA0001188718000000121
While the true distribution p (X) of X is fixed, the above problem can be transformed into finding q by an argument lower Bound (ELBO)*(Z) a maximization satisfying the following formula:
Figure BDA0001188718000000122
while
Figure BDA0001188718000000123
Therefore, the optimization problem can be converted into finding the target variational distribution function q corresponding to p (Z | X)*(Z) is such that the above formula takes the maximum value.
p (X | Z) is an expression based on the IRT model:
Figure BDA0001188718000000124
since q (Z) is with respect to the hyper-parameter tauα、τβAnd τθSo that L (q) is also related to the hyper-parameter τα、τβAnd τθSo that the hyper-parameter satisfies the maximization of L (q), i.e. the optimization problem of L (q) is translated into a function with respect to τα、τβAnd τθAnd the undetermined coefficient is solved, so that the super parameter is estimated and the parameter value of the super parameter is obtained.
Will estimate the obtained tauα、τβAnd τθthe updated q (Z) is replaced by the updated q (Z) which is the target variation distribution function q (α) p (β) p (θ)*(Z)。
the method comprises the following steps of sampling parameters to be estimated based on updated q (Z) to obtain estimation of the parameters to be estimated, wherein the specific sampling mode is not limited in the embodiment of the invention.
And finally, substituting the estimation result of the parameter to be estimated into the IRT model to obtain a second IRT prediction model.
The second IRT prediction model can reduce the influence of excessive solidification of the prior estimation of the parameter to be estimated on the estimation result, and effectively improves the estimation accuracy.
Further, after obtaining the first IRT prediction model and the second IRT prediction model, the method may further include: and assuming that the evolution of the learning ability of the answerer meets the wiener process, and updating the first IRT prediction model and the second IRT prediction model. Determining a first current learning capacity of the current answerer according to historical answer information and the updated first IRT prediction model; and determining the second current learning capacity of the current answerer according to the historical answer information and the updated second IRT prediction model. Predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the learning capability information and the first question information of the candidate question to obtain a first IRT prediction result, wherein the first IRT prediction result comprises: and predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the first current learning capability and the first question information of the candidate question to obtain a first IRT prediction result. Predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the learning capability information and the second question information of the candidate question to obtain a second IRT prediction result, wherein the second IRT prediction result comprises: and predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the second current learning capability and the second question information of the candidate question to obtain a second IRT prediction result.
Specifically, the change in the learning ability of the answerer is a process that evolves over time, and therefore, the embodiment of the present invention considers this factor and performs further optimization. It can be assumed that the evolution of the learning ability of the answerer satisfies the wiener process as follows:
Figure BDA0001188718000000131
where γ is the smoothed a priori assumption parameter of the wiener process, θt′+τTo the current learning ability of the answerer, thetat′The t-t 'represents the time interval between two question making for the learning ability of the last question making time t' of the answerer.
Adding the above assumptions into the prediction model (the first IRT prediction model or the second IRT prediction model), that is, at any time t, for a time point t' before any t, updating the prediction model to obtain an updated prediction model as follows:
Figure BDA0001188718000000141
wherein the content of the first and second substances,
Figure BDA0001188718000000142
Figure BDA0001188718000000143
Figure BDA0001188718000000144
represents the correction resolution, theta, of the topic j at time ti,tIndicating the current learning ability, X, of the answerer ii,j,t′Showing the wrong question setting condition of the question j at the time t' of the question i, X i,j,t′1 denotes the answerer i to make a question pair j at time t'.
And then determining the current learning ability of the current answerer according to the historical answer data and the updated prediction model. Specifically, the learning ability of the current answerer at the current moment can be estimated by using the updated prediction model in a maximum posterior probability estimation mode, and the learning ability of the answerer can be smoothed by using the method, so that the prediction precision can be further improved.
And step 140, tracking the DKT prediction model and the historical answer information based on the deep learning knowledge to predict the probability of the current answer to the candidate question, so as to obtain a DKT prediction result.
Illustratively, the DKT prediction model is a prediction model based on a DKT network model, historical answer information of a current answerer is input into the model, and output data is a prediction result of a current answerer about a question making situation of an unprocessed question in a target question bank, and a probability of making a question can be obtained. The intermediate layer in the DKT network may be any one of a Recurrent Neural Networks (RNN), a Long Short Term Memory (LSTM) network, and a threshold recursive Unit (GRU) network.
Preferably, the step may specifically include: clustering the answer information samples based on the number of the answers of the answerers to obtain a plurality of training sample subsets; training the DKT network by sequentially utilizing each training sample subset in the plurality of training sample subsets in an iterative mode to obtain a DKT prediction model; and predicting the probability of the current answer to the candidate question based on the DKT prediction model and the historical answer information to obtain a DKT prediction result.
Specifically, the DKT network is trained in an iterative manner, and the process of obtaining the DKT prediction model is as follows: training an initial DKT network by adopting a first training sample subset to obtain a DKT network intermediate model corresponding to the first training sample subset; for each training sample subset from a second training sample subset to a penultimate training sample subset, training a DKT network intermediate model corresponding to a previous training sample subset by using a current training sample subset to obtain a DKT network intermediate model corresponding to the current training sample subset; and training the DKT network intermediate model corresponding to the penultimate training sample subset by adopting the last training sample subset to obtain a DKT prediction model.
The method has the advantages that overfitting errors caused by severe fluctuation of the number of questions to be made by different answerers to the DKT model can be reduced, and prediction precision and accuracy of the DKT network model are improved.
Further, clustering the answer information samples based on the number of questions to be made by the answerers to obtain a plurality of training sample subsets, including: clustering the answer information samples based on the number of the answers of the answerers to obtain a plurality of answer information sample subsets; and for each answer information sample subset, converting the answer information corresponding to each answer in the current answer information sample subset into a unique heat vector based on the question and the knowledge point to obtain a training sample subset corresponding to the current answer information sample subset. When vectorization is carried out, more information about each topic can be obtained through the added knowledge point information, and the model prediction accuracy can be further improved. The converting the answer information corresponding to each answer in the current answer information sample subset into a unique heat vector based on the question and the knowledge point to obtain a training sample subset corresponding to the current answer information sample subset may include: respectively generating a first unique heat vector based on question answering conditions and a second unique heat vector based on knowledge point answering conditions of the current answering person according to answer information corresponding to each answering person in the current answer information sample subset; performing direct summation operation on the first unique heat vector and the second unique heat vector to obtain a unique heat vector based on the question and the knowledge point of the current answerer; and summarizing the independent heat vectors based on the questions and the knowledge points, which correspond to all the respondents contained in the current answer information sample subset, into a training sample subset corresponding to the current answer information sample subset.
Further, summarizing the unique heat vectors based on the questions and the knowledge points, which correspond to all the respondents contained in the current answer information sample subset, into the training sample subset corresponding to the current answer information sample subset, includes: performing compression reconstruction on the obtained one-hot vector by utilizing the sparsity of the obtained one-hot vector; and summarizing the compressed and reconstructed vectors corresponding to all the answerers contained in the current answer information sample subset into a training sample subset corresponding to the current answer information sample subset. The advantage of performing the compression reconstruction is that the situations that the training efficiency is affected and the machine memory is insufficient due to the overlarge vector length can be avoided.
Further, before each of the training sample subsets is used to train the DKT network in an iterative manner to obtain the DKT prediction model, the method may further include: and carrying out tuning treatment on the DKT network, wherein the tuning content corresponding to the tuning treatment comprises at least one of the network layer number, the coupling relation between networks, the type of each layer of network, the selection of an activation function layer, the selection of an objective function, a truncation threshold of a gradient, a learning coefficient of self-adaptive optimization and a random initialization scheme of a network weight.
Further, before each of the training sample subsets is used to train the DKT network in an iterative manner to obtain the DKT prediction model, the method may further include: DKT networks are improved based on Bucket mechanisms or dynamic neural network mechanisms.
And 150, combining the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
Illustratively, the IRT prediction result and the DKT prediction result can be subjected to nonlinear coupling processing to obtain the prediction probability of the current answerer on the candidate question.
Specifically, the IRT prediction result and the DKT prediction result may be combined based on the following formula to obtain the prediction probability of the current answerer for the candidate question:
PTemp＝ωMML*WMML*PMML+ωVI*WVI*PVI+ωDKT*WDKT*PDKT
Figure BDA0001188718000000171
Figure BDA0001188718000000172
Figure BDA0001188718000000173
wherein, PTempIs a predicted probability; omegaMMLConstant factor, W, for first IRT predictionMMLWeight coefficient, P, for the first IRT prediction resultMMLA probability value corresponding to the first IRT prediction result; omegaVIConstant factor, W, for second IRT predictionVIFor the second IRT prediction result, corresponding weight coefficient, PVIThe probability value corresponding to the second IRT prediction result is obtained; omegaDKTConstant factor, W, corresponding to DKT predictionDKTWeight coefficient, P, corresponding to DKT predictionDKTAnd the probability value is corresponding to the DKT prediction result.
In this embodiment, for each preThe value of the constant factor corresponding to the measurement result is not limited, and can be selected by a person skilled in the art according to actual requirements. Preferably, ω isMML＝ωVI＝0.25，ωDKT＝0.5。
Preferably, the prediction probability may be further smoothed to obtain a target prediction probability. Specifically, the prediction probability may be smoothed by the following formula to obtain a target prediction probability:
Figure BDA0001188718000000174
Figure BDA0001188718000000175
wherein, PFinalPredicting a probability for the target; s is a preset constant; omega is a first smoothing factor, and omega belongs to [0,1 ]](ii) a Lambda is a second smoothing factor, lambda belongs to (— infinity, 0), and the specific values of omega and lambda depend on the quality of the topic model and the positive and negative skewness values of the data; pDummyThe intermediate prediction probability;
m is the number of the questions to be made of the current answerer in the target question bank, and N is the total number of the questions to be made of the current answerer in the target question bank; or M is the number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, N is the total number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, and the specified knowledge point is the knowledge point to which the candidate question belongs. The specific values of M and N can be determined according to whether the answerer has made the question in the designated knowledge point.
In this embodiment, specific values of the preset constant s are not limited, and preferably, s is 0.5.
That is, the above formula is preferably:
Figure BDA0001188718000000181
the above formula can be called Dummy model when PTempA lower confidence around 0.5 indicates that the model prediction isOr the variance is larger, so the smoothing processing needs to be carried out, and the prediction result is more accurate.
The information prediction method provided by the embodiment of the invention predicts the probability of the candidate questions based on the IRT prediction model and the learning ability information of the current answerer for the candidate questions in the target question bank to obtain the IRT prediction result, predicts the probability of the candidate questions based on the DKT prediction model and the historical answer information of the current answerer to obtain the DKT prediction result, and finally combines the IRT prediction result and the DKT prediction result to obtain the prediction probability of the candidate questions for the current answerer. By adopting the technical scheme, the prediction results of the two prediction models are integrated to predict the probability of the candidate questions of the answerer, so that the prediction accuracy can be effectively improved.
Example two
Fig. 2 is a schematic flow chart of an information prediction method according to a second embodiment of the present invention, which is optimized based on the second embodiment.
Correspondingly, the method of the embodiment comprises the following steps:
and step 210, obtaining historical answer information of the current answer user about the target question bank and learning ability information of the current answer user.
Step 220, candidate topics are determined from the target topic library.
And step 230, predicting the probability of the current answerer for the candidate question based on the IRT prediction model and the learning ability information to obtain an IRT prediction result.
And step 240, tracking the DKT prediction model and the historical answer information based on the deep learning knowledge to predict the probability of the current answer to the candidate question, so as to obtain a DKT prediction result.
And step 250, combining the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
And step 260, smoothing the prediction probability to obtain the target prediction probability.
And step 270, when the target prediction probability meets a preset condition, pushing the candidate question to the current answer.
For example, the preset condition may be determined according to a default setting of the adaptive learning system, or may be set by the answerer according to the self condition. For example, the preset condition may be that the determined probability is within a preset numerical range, assuming that the range is 0.5-0.8, such as for candidate topic C, when the determined probability is 0.6, then the topic C is pushed to the current answerer.
Preferably, the step may specifically include:
defining entropy values of the candidate topics as:
H＝-PFinallogPFinal-(1-PFinal)log(1-PFinal)
wherein, PFinalH is the Entropy value (Information entry) of the candidate topic when the probability is determined.
When P is presentFinalAnd when the H value is larger than the preset value, pushing the candidate question to the current answer.
It can be understood that according to the maximum entropy principle, the larger the entropy value of a candidate question, the more information quantity can be obtained by a question answering person practicing the question, so that when the H value is larger than a certain value, the candidate question is pushed to the current question answering person.
Fig. 3 is a schematic diagram of a question pushing process according to a second embodiment of the present invention, as shown in fig. 3, taking an answerer as a student as an example, a rear knowledge point, a front knowledge point or a parallel knowledge point can be selected according to a knowledge point map and the learning ability of the student, and a question pushed to the student is selected. As can be seen from the answer prediction graph of the student 1001 in fig. 3, when the answer accuracy of the student is predicted to be 50% (the learning ability of the student is equal to the difficulty of the question), the entropy of the information is the largest, and the information amount of the question is the largest for the student, so that the student 1001 pushes the question 2023 under the knowledge point of the "exclamation sentence" in this example.
According to the information prediction method provided by the embodiment of the invention, after the probability of the candidate question of the answerer is predicted, the appropriate question is quickly and accurately selected and pushed to the answerer for answering based on the predicted probability and the current learning capacity of the current answerer, so that the self-adaptive learning system has pertinence and individuation, the learning effect of the answerer is maximized, and the inefficient situations that the answerer cannot do or cannot harvest due to too many simple questions or directly do difficult problems are avoided.
Fig. 4a is a first comparison diagram of the technical solution of the embodiment of the present invention and the prior art, and fig. 4b is a second comparison diagram of the technical solution of the embodiment of the present invention and the prior art. Fig. 4a is a graph comparing accuracy of a classical IRT model (Knewton HIRT) used by Knewton corporation and a DKT model (Stanford RNN) proposed by Stanford university, which are obtained based on assistent open source data (assistent 2009-2010 "skin builder" dataset), and a lean learning model formed by the technical solution of the embodiment of the present invention; fig. 4b is a diagram showing the accuracy comparison between the classical IRT model used by Knewton corporation and the DKT model proposed by stanford university, which are obtained based on scholar monarch live course data, and the lean learning model formed by the technical solution of the embodiment of the present invention. As shown in fig. 4a and 4b, the lean learning model provided by the embodiment of the present invention improves the effect to different degrees no matter on the public data source or the data source of the scholar man, and has better accuracy while considering the operation efficiency.
EXAMPLE III
Fig. 5 is a block diagram of an information prediction apparatus provided in a third embodiment of the present invention, where the apparatus may be implemented by software and/or hardware, and may be generally integrated in a terminal in an adaptive learning system, where the terminal may be a terminal such as a personal computer or a server, or a mobile terminal such as a tablet computer or a smart phone, and the embodiment of the present invention is not limited in particular. As shown in fig. 5, the apparatus includes an information acquisition module 31, a candidate topic determination module 32, a first prediction module 33, a second prediction module 34, and a prediction result processing module 35.
The information acquiring module 31 is configured to acquire historical answer information of a current answer user about a target question bank and learning ability information of the current answer user;
a candidate question determining module 32, configured to determine candidate questions from the target question bank;
the first prediction module 33 is configured to predict, based on a project reflection theory IRT prediction model and the learning capability information, a probability that the current answerer performs on the candidate question, and obtain an IRT prediction result;
a second prediction module 34, configured to track a DKT prediction model based on deep learning knowledge and predict, based on the historical answer information, a probability that the current answer person makes the candidate question, so as to obtain a DKT prediction result;
and the prediction result processing module 35 is configured to perform merging processing on the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
The information prediction device provided by the embodiment of the invention synthesizes the prediction results of the two prediction models to predict the probability of the candidate questions made by the answerer, and can effectively improve the prediction accuracy.
On the basis of the above embodiment, the apparatus further includes: an answer sample obtaining module, configured to obtain answer information samples of a preset number of answer users regarding the target question bank before predicting, based on the IRT model and the learning ability information, a probability that the current answer user performs on the candidate question and obtaining an IRT prediction result; the frequency weight determining module is used for determining the frequency weight of the corresponding question in the target question bank which is paired by all the respondents with the preset learning ability on the basis of the assumption that the answering condition of the question made by each responder is an independent event; the frequency weight substituting module is used for substituting the frequency weight into a preset estimation model to obtain a target estimation model, wherein the preset estimation model is a maximum posterior probability estimation model based on an IRT (inverse Fourier transform) model and a Marginal Maximum Likelihood (MML) estimation method; the information estimation module is used for training the target estimation model by utilizing an MML estimation method according to the answer sample information so as to estimate first question information of each question in the target question bank, wherein the first question information comprises a first distinguishing degree and a first difficulty; and the first prediction model establishing module is used for establishing a first IRT prediction model according to the trained target estimation model. The first prediction module is to: and predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the learning capability information and the first question information of the candidate question to obtain a first IRT prediction result.
On the basis of the above embodiment, the apparatus further includes: the Bayesian network model building module is used for building a Bayesian network model by taking the learning capacity of the respondents, the discrimination of questions and the difficulty of the questions in the IRT model as parameters to be estimated after acquiring answer information samples of a preset number of respondents about the target question bank, wherein the parameters to be estimated meet preset prior distribution containing the hyper-parameters; the hyperparameter estimation module is used for determining a variation distribution function corresponding to a target function by adopting a variation inference method, and estimating the hyperparameter based on the Bayesian network model and the answer information sample by taking the minimum degree of closeness of the target function and the variation distribution function as a principle to obtain a parameter value of the hyperparameter, wherein the target function is a posterior estimation function about the parameter to be estimated based on the answer information sample; the function updating module is used for updating the variation distribution function according to the obtained parameter values of the hyper-parameters; the parameter estimation module to be estimated is used for sampling the parameter to be estimated based on the updated variation distribution function to obtain the estimation of the parameter to be estimated; and the first prediction model establishing module is used for establishing a second IRT prediction model according to the estimation result of the parameter to be estimated, wherein the estimation result comprises second topic information of each topic in the target topic library, and the second topic information comprises second discrimination and second difficulty. The first prediction module is further to: and predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the learning capability information and the second question information of the candidate question to obtain a second IRT prediction result.
On the basis of the foregoing embodiment, the second prediction module is configured to: clustering the answer information samples based on the number of the answers of the answerers to obtain a plurality of training sample subsets;
training the DKT network by sequentially utilizing each training sample subset in the plurality of training sample subsets in an iterative mode to obtain a DKT prediction model;
and predicting the probability of the current answer to the candidate question based on the DKT prediction model and the historical answer information to obtain a DKT prediction result.
On the basis of the foregoing embodiment, the prediction result processing module is configured to:
and combining the IRT prediction result and the DKT prediction result based on the following formula to obtain the prediction probability of the current answerer on the candidate question:
PTemp＝ωMML*WMML*PMML+ωVI*WVI*PVI+ωDKT*WDKT*PDKT
Figure BDA0001188718000000231
Figure BDA0001188718000000232
Figure BDA0001188718000000233
wherein, PTempIs a predicted probability; omegaMMLConstant factor, W, for first IRT predictionMMLWeight coefficient, P, for the first IRT prediction resultMMLA probability value corresponding to the first IRT prediction result; omegaVIConstant factor, W, for second IRT predictionVIFor the second IRT prediction result, corresponding weight coefficient, PVIThe probability value corresponding to the second IRT prediction result is obtained; omegaDKTConstant factor, W, corresponding to DKT predictionDKTWeight coefficient, P, corresponding to DKT predictionDKTAnd the probability value is corresponding to the DKT prediction result.
On the basis of the above embodiment, the apparatus further includes:
a smoothing processing module, configured to perform nonlinear coupling processing on the IRT prediction result and the DKT prediction result to obtain a prediction probability of the candidate question for the current answerer, and then perform smoothing processing on the prediction probability based on the following formula to obtain a target prediction probability:
Figure BDA0001188718000000241
Figure BDA0001188718000000242
wherein, PFinalPredicting a probability for the target; s is a preset constant; omega is a first smoothing factor, and omega belongs to [0,1 ]](ii) a λ is the second smoothing factor, λ ∈ (— ∞, 0); pDummyThe intermediate prediction probability;
m is the number of the questions to be made of the current answerer in the target question bank, and N is the total number of the questions to be made of the current answerer in the target question bank; or M is the number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, N is the total number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, and the specified knowledge point is the knowledge point to which the candidate question belongs.
On the basis of the above embodiment, the device further includes a question pushing module, configured to, after the prediction probability is subjected to smoothing processing to obtain a target prediction probability, push the candidate question to the current answerer when the target prediction probability meets a preset condition.
On the basis of the above embodiment, the apparatus further includes: the prediction model updating module is used for assuming that the evolution of the learning capacity of the answerer meets the wiener process after the first IRT prediction model and the second IRT prediction model are obtained, and updating the first IRT prediction model and the second IRT prediction model; the first ability determining module is used for determining the first current learning ability of the current answerer according to the historical answer information and the updated first IRT prediction model; and the second ability determining module is used for determining the second current learning ability of the current answerer according to the historical answer information and the updated first IRT prediction model. The first prediction module is to: predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the first current learning capacity and the first question information of the candidate question to obtain a first IRT prediction result; and predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the second current learning capability and the second question information of the candidate question to obtain a second IRT prediction result.
The information prediction device provided in the above embodiment can execute the information prediction method provided in any embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the method. The technical details not described in detail in the above embodiments may be referred to the information prediction method provided in any embodiment of the present invention.
It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims (20)

1. An information prediction method, comprising:
acquiring historical answer information of a current answer user about a target question bank and learning capacity information of the current answer user;
determining candidate questions from the target question bank;
predicting the probability of the current answerer for the candidate question based on a project reflection theory IRT prediction model and the learning ability information to obtain an IRT prediction result;
predicting the probability of the current answerer for the candidate question based on a deep learning knowledge tracking DKT prediction model and the historical answer information to obtain a DKT prediction result;
and combining the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
2. The method of claim 1, further comprising, before predicting the probability of the current answerer making the candidate question based on an IRT model and the learning capability information, and obtaining an IRT prediction result:
acquiring answer information samples of a preset number of answer users about the target question bank;
determining the frequency weight of the corresponding question in the target question bank which is paired by all the answering persons with preset learning capacity based on the assumption that the answering condition of the question made by each answering person is an independent event;
substituting the frequency weight into a preset estimation model to obtain a target estimation model, wherein the preset estimation model is a maximum posterior probability estimation model based on an IRT (inverse Fourier transform) model and a Marginal Maximum Likelihood (MML) estimation method;
training the target estimation model by using an MML estimation method according to the answer sample information to estimate first question information of each question in the target question bank, wherein the first question information comprises a first discrimination and a first difficulty;
establishing a first IRT prediction model according to the trained target estimation model;
predicting the probability of the current answerer for the candidate question based on the project reflection theory IRT model and the learning ability information to obtain an IRT prediction result, wherein the IRT prediction result comprises the following steps:
and predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the learning capability information and the first question information of the candidate question to obtain a first IRT prediction result.
3. The method of claim 2, wherein after obtaining samples of answer information of a preset number of respondents about the target question bank, the method further comprises:
constructing a Bayesian network model by taking the learning ability of an answerer, the discrimination of questions and the difficulty of the questions in an IRT model as parameters to be estimated, wherein the parameters to be estimated meet the preset prior distribution containing hyper-parameters;
determining a variation distribution function corresponding to a target function by adopting a variation inference method, and estimating the hyper-parameter based on the Bayesian network model and the answer information sample by taking the minimum degree of closeness of the target function and the variation distribution function as a principle to obtain a parameter value of the hyper-parameter, wherein the target function is a posterior estimation function about the parameter to be estimated based on the answer information sample;
updating the variation distribution function according to the obtained parameter values of the hyperparameters;
sampling the parameter to be estimated based on the updated variational distribution function to obtain the estimation of the parameter to be estimated;
establishing a second IRT prediction model according to an estimation result of the parameter to be estimated, wherein the estimation result comprises second topic information of each topic in the target topic library, and the second topic information comprises second discrimination and second difficulty;
predicting the probability of the current answerer for the candidate question based on the project reflection theory IRT model and the learning ability information to obtain an IRT prediction result, and further comprising:
and predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the learning capability information and the second question information of the candidate question to obtain a second IRT prediction result.
4. The method of claim 3, wherein predicting the probability of the current responder making the candidate question based on a DKT prediction model and the historical answer information to obtain a DKT prediction result comprises:
clustering the answer information samples based on the number of the answers of the answerers to obtain a plurality of training sample subsets;
training the DKT network by sequentially utilizing each training sample subset in the plurality of training sample subsets in an iterative mode to obtain a DKT prediction model;
and predicting the probability of the current answer to the candidate question based on the DKT prediction model and the historical answer information to obtain a DKT prediction result.
5. The method according to claim 4, wherein the combining the IRT prediction result and the IRT prediction result to obtain the prediction probability of the current answerer for the candidate question comprises:
and carrying out nonlinear coupling processing on the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
6. The method of claim 5, wherein performing nonlinear coupling processing on the IRT prediction result and the DKTT prediction result to obtain a prediction probability of the current answerer for the candidate question, comprises:
and combining the IRT prediction result and the DKT prediction result based on the following formula to obtain the prediction probability of the current answerer on the candidate question:
PTemp＝ωMML*WMML*PMML+ωVI*WVI*PVI+ωDKT*WDKT*PDKT
Figure FDA0001188717990000031
Figure FDA0001188717990000032
Figure FDA0001188717990000041
wherein, PTempIs a predicted probability; omegaMMLConstant factor, W, for first IRT predictionMMLWeight coefficient, P, for the first IRT prediction resultMMLA probability value corresponding to the first IRT prediction result; omegaVIConstant factor, W, for second IRT predictionVIFor the second IRT prediction result, corresponding weight coefficient, PVIThe probability value corresponding to the second IRT prediction result is obtained; omegaDKTConstant factor, W, corresponding to DKT predictionDKTWeight coefficient, P, corresponding to DKT predictionDKTAnd the probability value is corresponding to the DKT prediction result.
7. The method of claim 6, wherein after performing nonlinear coupling processing on the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer for the candidate question, the method further comprises:
and smoothing the prediction probability to obtain a target prediction probability.
8. The method of claim 7, wherein smoothing the prediction probability to obtain a target prediction probability comprises:
smoothing the prediction probability based on the following formula to obtain a target prediction probability:
Figure FDA0001188717990000042
Figure FDA0001188717990000043
wherein, PFinalPredicting a probability for the target; s is a preset constant; omega is a first smoothing factor, and omega belongs to [0,1 ]](ii) a λ is the second smoothingFactor, λ ∈ (— ∞, 0); pDummyThe intermediate prediction probability;
m is the number of the questions to be made of the current answerer in the target question bank, and N is the total number of the questions to be made of the current answerer in the target question bank; or M is the number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, N is the total number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, and the specified knowledge point is the knowledge point to which the candidate question belongs.
9. The method of claim 8, wherein s is 0.5.
10. The method of claim 9, further comprising, after smoothing the prediction probability to obtain a target prediction probability:
and when the target prediction probability meets a preset condition, pushing the candidate question to the current answer.
11. The method of claim 10, wherein when the target prediction probability satisfies a preset condition, pushing the candidate question to the current answerer comprises:
defining entropy values H of the candidate questions as:
H＝-PFinallog PFinal-(1-PFinal)log(1-PFinal)
when P is presentFinalAnd when the H value is larger than a preset value, pushing the candidate question to the current answer.
12. The method of claim 3, after obtaining the first IRT prediction model and the second IRT prediction model, further comprising:
assuming that the evolution of the learning ability of the answerer meets the wiener process, and updating the first IRT prediction model and the second IRT prediction model;
determining a first current learning capacity of the current answerer according to the historical answer information and the updated first IRT prediction model;
determining a second current learning capacity of the current answerer according to the historical answer information and the updated second IRT prediction model;
predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the learning capability information and the first question information of the candidate question to obtain a first IRT prediction result, wherein the first IRT prediction result comprises:
predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the first current learning capacity and the first question information of the candidate question to obtain a first IRT prediction result;
predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the learning capability information and the second question information of the candidate question to obtain a second IRT prediction result, wherein the second IRT prediction result comprises:
and predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the second current learning capability and the second question information of the candidate question to obtain a second IRT prediction result.
13. An information prediction apparatus, comprising:
the information acquisition module is used for acquiring historical answer information of a current answer user about a target question bank and learning capacity information of the current answer user;
the candidate question determining module is used for determining candidate questions from the target question bank;
the first prediction module is used for predicting the probability of the current answerer for the candidate question based on a project reflection theory IRT prediction model and the learning capability information to obtain an IRT prediction result;
the second prediction module is used for predicting the probability of the current answer to the candidate question based on a deep learning knowledge tracking DKT prediction model and the historical answer information to obtain a DKT prediction result;
and the prediction result processing module is used for carrying out merging processing on the IRT prediction result and the DKT prediction result to obtain the prediction probability of the current answerer on the candidate question.
14. The apparatus of claim 13, further comprising:
an answer sample obtaining module, configured to obtain answer information samples of a preset number of answer users regarding the target question bank before predicting, based on the IRT model and the learning ability information, a probability that the current answer user performs on the candidate question and obtaining an IRT prediction result;
the frequency weight determining module is used for determining the frequency weight of the corresponding question in the target question bank which is paired by all the respondents with the preset learning ability on the basis of the assumption that the answering condition of the question made by each responder is an independent event;
the frequency weight substituting module is used for substituting the frequency weight into a preset estimation model to obtain a target estimation model, wherein the preset estimation model is a maximum posterior probability estimation model based on an IRT (inverse Fourier transform) model and a Marginal Maximum Likelihood (MML) estimation method;
the information estimation module is used for training the target estimation model by utilizing an MML estimation method according to the answer sample information so as to estimate first question information of each question in the target question bank, wherein the first question information comprises a first distinguishing degree and a first difficulty;
the first prediction model establishing module is used for establishing a first IRT prediction model according to the trained target estimation model;
the first prediction module is to:
and predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the learning capability information and the first question information of the candidate question to obtain a first IRT prediction result.
15. The apparatus of claim 14, further comprising:
the Bayesian network model building module is used for building a Bayesian network model by taking the learning capacity of the respondents, the discrimination of questions and the difficulty of the questions in the IRT model as parameters to be estimated after acquiring answer information samples of a preset number of respondents about the target question bank, wherein the parameters to be estimated meet preset prior distribution containing the hyper-parameters;
the hyperparameter estimation module is used for determining a variation distribution function corresponding to a target function by adopting a variation inference method, and estimating the hyperparameter based on the Bayesian network model and the answer information sample by taking the minimum degree of closeness of the target function and the variation distribution function as a principle to obtain a parameter value of the hyperparameter, wherein the target function is a posterior estimation function about the parameter to be estimated based on the answer information sample;
the function updating module is used for updating the variation distribution function according to the obtained parameter values of the hyperparameter;
the parameter estimation module to be estimated is used for sampling the parameter to be estimated based on the updated variation distribution function to obtain the estimation of the parameter to be estimated;
the second prediction model establishing module is used for establishing a second IRT prediction model according to the estimation result of the parameter to be estimated, wherein the estimation result comprises second topic information of each topic in the target topic library, and the second topic information comprises second discrimination and second difficulty;
the first prediction module is further to:
and predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the learning capability information and the second question information of the candidate question to obtain a second IRT prediction result.
16. The apparatus of claim 15, wherein the second prediction module is configured to:
clustering the answer information samples based on the number of the answers of the answerers to obtain a plurality of training sample subsets;
training the DKT network by sequentially utilizing each training sample subset in the plurality of training sample subsets in an iterative mode to obtain a DKT prediction model;
and predicting the probability of the current answer to the candidate question based on the DKT prediction model and the historical answer information to obtain a DKT prediction result.
17. The apparatus of claim 16, wherein the prediction processing module is configured to:
and combining the IRT prediction result and the DKT prediction result based on the following formula to obtain the prediction probability of the current answerer on the candidate question:
PTemp＝ωMML*WMML*PMML+ωVI*WVI*PVI+ωDKT*WDKT*PDKT
Figure FDA0001188717990000081
Figure FDA0001188717990000091
Figure FDA0001188717990000092
wherein, PTempIs a predicted probability; omegaMMLConstant factor, W, for first IRT predictionMMLWeight coefficient, P, for the first IRT prediction resultMMLA probability value corresponding to the first IRT prediction result; omegaVIConstant factor, W, for second IRT predictionVIFor the second IRT prediction result, corresponding weight coefficient, PVIThe probability value corresponding to the second IRT prediction result is obtained; omegaDKTConstant factor, W, corresponding to DKT predictionDKTWeight coefficient, P, corresponding to DKT predictionDKTMapping for DKT prediction resultsThe probability value of (2).
18. The apparatus of claim 17, further comprising:
a smoothing processing module, configured to perform nonlinear coupling processing on the IRT prediction result and the DKT prediction result to obtain a prediction probability of the candidate question for the current answerer, and then perform smoothing processing on the prediction probability based on the following formula to obtain a target prediction probability:
Figure FDA0001188717990000093
Figure FDA0001188717990000094
wherein, PFinalPredicting a probability for the target; s is a preset constant; omega is a first smoothing factor, and omega belongs to [0,1 ]](ii) a λ is the second smoothing factor, λ ∈ (— ∞, 0); pDummyThe intermediate prediction probability;
m is the number of the questions to be made of the current answerer in the target question bank, and N is the total number of the questions to be made of the current answerer in the target question bank; or M is the number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, N is the total number of the questions to be asked of the current answerer under the specified knowledge point in the target question bank, and the specified knowledge point is the knowledge point to which the candidate question belongs.
19. The apparatus of claim 18, further comprising:
and the question pushing module is used for pushing the candidate question to the current answerer when the target prediction probability meets a preset condition after the prediction probability is subjected to smoothing processing to obtain the target prediction probability.
20. The apparatus of claim 15, further comprising:
the prediction model updating module is used for assuming that the evolution of the learning capacity of the answerer meets the wiener process after the first IRT prediction model and the second IRT prediction model are obtained, and updating the first IRT prediction model and the second IRT prediction model;
the first ability determining module is used for determining the first current learning ability of the current answerer according to the historical answer information and the updated first IRT prediction model;
the second ability determining module is used for determining a second current learning ability of the current answerer according to the historical answer information and the updated second IRT prediction model;
the first prediction module is to:
predicting the probability of the current answerer for the candidate question based on the first IRT prediction model, the first current learning capacity and the first question information of the candidate question to obtain a first IRT prediction result;
and predicting the probability of the current answerer for the candidate question based on the second IRT prediction model, the second current learning capability and the second question information of the candidate question to obtain a second IRT prediction result.
CN201611199287.0A 2016-12-22 2016-12-22 Information prediction method and device Active CN108229718B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201611199287.0A CN108229718B (en) 2016-12-22 2016-12-22 Information prediction method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201611199287.0A CN108229718B (en) 2016-12-22 2016-12-22 Information prediction method and device

Publications (2)

Publication Number Publication Date
CN108229718A CN108229718A (en) 2018-06-29
CN108229718B true CN108229718B (en) 2020-06-02

Family

ID=62656238

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201611199287.0A Active CN108229718B (en) 2016-12-22 2016-12-22 Information prediction method and device

Country Status (1)

Country Link
CN (1) CN108229718B (en)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108985290A (en) * 2018-07-18 2018-12-11 杜艳平 A kind of intelligent check system for teachers ' teaching quality
CN109086939A (en) * 2018-08-14 2018-12-25 合肥知多少教育科技有限公司 A kind of input method and device doing topic result
US20200202226A1 (en) * 2018-12-20 2020-06-25 Fuji Xerox Co., Ltd. System and method for context based deep knowledge tracing
CN109685564B (en) * 2018-12-21 2021-02-19 苏州易泰勒电子科技有限公司 Offline consumer decision behavior prediction method based on project reaction theory
CN111353015B (en) * 2018-12-24 2024-03-15 阿里巴巴集团控股有限公司 Crowd-sourced question recommendation method, device, equipment and storage medium
CN109829583B (en) * 2019-01-31 2022-10-11 成都思晗科技股份有限公司 Mountain fire risk prediction method based on probability programming technology
CN110377814A (en) * 2019-05-31 2019-10-25 平安国际智慧城市科技股份有限公司 Topic recommended method, device and medium
CN110472060B (en) * 2019-07-05 2020-12-01 平安国际智慧城市科技股份有限公司 Topic pushing method and device, computer equipment and storage medium
CN112711704A (en) * 2019-10-25 2021-04-27 北京一起教育信息咨询有限责任公司 Exercise pushing method and device
CN111428021B (en) * 2020-06-05 2023-05-30 平安国际智慧城市科技股份有限公司 Text processing method, device, computer equipment and medium based on machine learning
CN112052828A (en) * 2020-09-23 2020-12-08 腾讯科技（深圳）有限公司 Learning ability determination method and device and storage medium
CN114418443A (en) * 2022-01-28 2022-04-29 华南师范大学 Test paper quality detection method, system, device and storage medium
CN114386716B (en) * 2022-02-16 2023-06-16 平安科技（深圳）有限公司 Answer sequence prediction method based on improved IRT structure, controller and storage medium

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104239969A (en) * 2014-09-04 2014-12-24 上海合煦信息科技有限公司 Evaluation and problem recommendation system for individualized education

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104239969A (en) * 2014-09-04 2014-12-24 上海合煦信息科技有限公司 Evaluation and problem recommendation system for individualized education

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Deep Knowledge Tracing;Chris Piech等;《arXiv：1506.05909v1》;20150619;第1-13页 *

Also Published As

Publication number Publication date
CN108229718A (en) 2018-06-29

Similar Documents

Publication Publication Date Title
CN108229718B (en) Information prediction method and device
CN109857835B (en) Self-adaptive network security knowledge evaluation method based on cognitive diagnosis theory
Bachrach et al. How to grade a test without knowing the answers---a Bayesian graphical model for adaptive crowdsourcing and aptitude testing
DiDonato et al. Social perception as induction and inference: An integrative model of intergroup differentiation, ingroup favoritism, and differential accuracy.
CN108228674B (en) DKT-based information processing method and device
Gan et al. Modeling learner’s dynamic knowledge construction procedure and cognitive item difficulty for knowledge tracing
US20210117842A1 (en) Systems and Methods for Training Generative Models Using Summary Statistics and Other Constraints
Davidson et al. Modeling accuracy as a function of response time with the generalized linear mixed effects model
CN108229688B (en) Information processing method and device based on IRT
CN114254208A (en) Identification method of weak knowledge points and planning method and device of learning path
Sälzer et al. Students’ skipping behavior on truancy items and (school) subjects and its relation to test performance in PISA 2012
Tian et al. Capacity tracing-enhanced course recommendation in MOOCs
Chen et al. Item Response Theory--A Statistical Framework for Educational and Psychological Measurement
Kuravsky et al. Probabilistic Modeling of a Testing Procedure
CN108229683B (en) Information processing method and device based on IRT
Wang et al. Varfa: A variational factor analysis framework for efficient bayesian learning analytics
Gowda et al. Improving Models of Slipping, Guessing, and Moment-By-Moment Learning with Estimates of Skill Difficulty.
CN109800880B (en) Self-adaptive learning feature extraction system based on dynamic learning style information and application
Baschera et al. Modeling engagement dynamics in spelling learning
Matayoshi et al. Studying retrieval practice in an intelligent tutoring system
CN115205072A (en) Cognitive diagnosis method for long-period evaluation
Kang et al. Personalized exercise recommendation via implicit skills
Kuravsky et al. Adaptive testing as a Markov process: models and their identification
Lee et al. Prescribing deep attentive score prediction attracts improved student engagement
CN113674116A (en) Student score prediction method and device based on fuzzy cloud cognitive diagnosis model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
TA01 Transfer of patent application right

Effective date of registration: 20180711

Address after: Room 3835, Building 3, 2528 Zhennan Road, Putuo District, Shanghai, 2003

Applicant after: Shanghai Ren Jing Information Technology Co., Ltd.

Address before: Room 2002, Building 5, 200 Guoding East Road, Yangpu District, Shanghai, 2003

Applicant before: Shanghai Qian Wan answer Cloud Computing Technology Co., Ltd.

TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20181011

Address after: 100041 B-0035, 2 floor, 3 building, 30 Shixing street, Shijingshan District, Beijing.

Applicant after: BEIJING ZIJIE TIAODONG NETWORK TECHNOLOGY CO., LTD.

Address before: Room 3835, Building 3, 2528 Zhennan Road, Putuo District, Shanghai, 2003

Applicant before: Shanghai Ren Jing Information Technology Co., Ltd.

TA01 Transfer of patent application right
GR01 Patent grant
GR01 Patent grant